Assigning a specific category to cells from varying levels

[yeroslaviz]

Hi Massimo, and thanks again for all the help until now for this great tool.

I'm trying to run you pre-built model scGate_models_DB$mouse$HiTME$panDC with my data.

I do get results, but not sure how to interpret them (see snippet below).

I understand I can plot the purity of the classification for the model, but is there a way to assign a category to the cell as well?

A cell is classified as "Pure" only if was scored as "Pure" in all levels, but then I get only "Target" in the "scGate_multi" column. What does it mean? How can I give it a more precise classification (is it possible at all)?

E.g. the fifth row(C7_AAACCCACATGAATCC-1 ) is classified as Pure on all levels with the highest value in "Myeloid_UCell". How can I assign this class to this cell?

I know one can assign a class using the multi-class model, but also only(!) for those cells, that were unique to a specific model in the calculations. Is there a way to assign a class also to the others, let's say based on the highest score calculated?
Would this make sense?

Plotting the data I have with this model:

DimPlot(C7_G7_sctransform_scgate_panDC, group.by = "is.pure", label = T, repel = T, label.size = 3) + theme(aspect.ratio = 1) +
    ggtitle("C7_G7") 

results in:

For testing reason I have tried to subset the model and use only 'level2' categories, but still, I get one column of is.Pure, but also here I get only "Target" but nothing more specific. How do I know if the Target here was cDC2 or Tcells or DC3 or etc.?

I appreciate your help

thanks again
Assa

Addendum

When Running the multi-class model with scGate_models_DB$mouse$HiTME I get the follwoing distribution of cell per category

### Detected a total of 12 pure 'Bcell' cells (0.06% of total)

### Detected a total of 40 pure 'CD4T' cells (0.20% of total)

### Detected a total of 104 pure 'CD8T' cells (0.51% of total)
Warning, all cells were removed at level 1. Consider reviewing signatures or model layout...

### Detected a total of 0 pure 'Endothelial' cells (0.00% of total)
Warning, all cells were removed at level 1. Consider reviewing signatures or model layout...

### Detected a total of 0 pure 'Epithelial' cells (0.00% of total)
Warning, all cells were removed at level 1. Consider reviewing signatures or model layout...

### Detected a total of 0 pure 'Erythrocyte' cells (0.00% of total)
Warning, all cells were removed at level 1. Consider reviewing signatures or model layout...

### Detected a total of 0 pure 'Fibroblast' cells (0.00% of total)
Warning, all cells were removed at level 3. Consider reviewing signatures or model layout...

### Detected a total of 0 pure 'gdT' cells (0.00% of total)

### Detected a total of 79 pure 'Mast' cells (0.39% of total)

### Detected a total of 6528 pure 'MoMac' cells (32.03% of total)

### Detected a total of 18 pure 'Neutrophils' cells (0.09% of total)

### Detected a total of 34 pure 'NK' cells (0.17% of total)

### Detected a total of 10085 pure 'panDC' cells (49.48% of total)
Warning, all cells were removed at level 3. Consider reviewing signatures or model layout...

### Detected a total of 0 pure 'PlasmaCell' cells (0.00% of total)

and the following DimPlot

DimPlot(C7_G7_sctransform_scgate_multi.model, group.by = "CellOntology_name", 
        label = T, repel = T, label.size = 2, reduction = "tsne") +
  theme(aspect.ratio = 1) + 
  ggtitle("scGate annotation") 

As my goal is to identify dendritic cells, would it make sense to extract those cell that are pure dendritic and recalculate the scores? Would this. help achieve better classification?

[mass-a]

Hi Assa,
perhaps this is not sufficiently clear in the documentation, but scGate can be run in two modes:

1. A single scGate model, as in your example for the panDC model. Using a single model, scGate will purify the target population (i.e. dendritic cells). This is what the "Target" metadata indicates.
2. A list of scGate models, as in your second example. In this case, scGate acts as a multi-class classifier, and combines the "Target" predictions for each of the models in the list to produce a cell label (scGate_multi metadata).

I understand your goal is to classify dendritic cell subtypes? If this is the case, you would have to specify a model for each DC subtype, and put them into a list. You could build these models from the existing panDC model, adding (for each model) an additional level with the signature of the DC subtype of interest. Does it make sense?

In our experience, marker-based classification works well up to a certain level of resolution. For high-resolution classification of subtypes, we tend to combine scGate with reference-based analysis. For example, this CASE STUDY uses ProjecTILs and a DC reference map to classify DC subtypes. In this approach scGate is used internally to first purify DCs with the panDC model.

Best
-massimo